Create Readme_BUS_UCLM.md

BUS-UCLM/Readme_BUS_UCLM.md

@@ -0,0 +1,140 @@
+# BUS-UCLM (Breast Ultrasound Lesion Segmentation Dataset) — packaged in `saiteja33/MedThink-Seg/BUS-UCLM`
+
+This README documents:
+1) **The original BUS-UCLM dataset** (what it contains, image sizes, labels, and original structure), and  
+2) **How BUS-UCLM is organized inside this Hugging Face repo** (`saiteja33/MedThink-Seg`, under the `BUS-UCLM/` directory).
+
+---
+
+## 1) Original BUS-UCLM dataset (source)
+
+**Primary reference (Data Descriptor paper):**  
+*“BUS-UCLM: Breast ultrasound lesion segmentation dataset”* (Scientific Data, 2025).  
+DOI: `10.1038/s41597-025-04562-3`
+
+**Original dataset record (Mendeley Data, Version 3):**  
+DOI: `10.17632/7fvgj4jsp7.3`
+
+### What images are in BUS-UCLM?
+- **Modality:** breast **ultrasound (BUS)** images (**PNG**)
+- **Clinical origin:** 2022–2023 scans from a hospital cohort (38 patients), acquired on a **Siemens ACUSON S2000** ultrasound system (18L6 HD probe) and later converted from DICOM to PNG.
+- **Image size:** **768 × 1024** pixels (fixed). Pixel spacing varies across the dataset (paper reports the most common as 0.0639205 mm/pixel).
+
+### Dataset size and counts
+- **Total images:** **683** (from **38** patients)
+  - **Normal:** 419
+  - **Benign:** 174
+  - **Malignant:** 90
+
+### Labels and annotations
+BUS-UCLM provides **pixel-wise segmentation masks** (stored as **RGB PNG** masks):
+- **Malignant lesion:** **red** = (255, 0, 0)
+- **Benign lesion:** **green** = (0, 255, 0)
+- **Background / normal tissue:** **black** = (0, 0, 0)
+
+### Original naming convention
+Files use:
+- **4 random letters** (anonymized patient identifier) + `_` + **sequence number** (image index within that patient), e.g.:
+  - `ALWI_000.png`, `ALWI_001.png`, …
+
+Typical layout:
+```
+BUS-UCLM/
+├─ images/              # ultrasound images (PNG)
+├─ masks/               # RGB segmentation masks (PNG)
+└─ <metadata>.csv       # image-level metadata (name/label/resolution/etc.)
+```
+
+**License (original):** CC BY 4.0 (per paper and Mendeley record)
+
+---
+
+## 2) BUS-UCLM inside `saiteja33/MedThink-Seg` (this Hugging Face repo)
+
+**Location in repo:** `BUS-UCLM/`  
+**Repo UI size (approx):** ~338 MB (as shown in the Hugging Face file browser)
+
+### What’s included here?
+This HF packaging organizes samples into:
+- **lesion cases** in `BUS-UCLM/images/` with masks in `BUS-UCLM/masks/`, and
+- **normal cases** in `BUS-UCLM/normal/images/` with masks in `BUS-UCLM/normal/masks/`.
+
+### Repo directory structure
+```
+BUS-UCLM/
+├─ images/                # lesion images (benign + malignant), PNG
+├─ masks/                 # corresponding masks for images/, PNG (same filenames)
+└─ normal/
+   ├─ images/             # normal images, PNG
+   └─ masks/              # corresponding masks for normal/images/, PNG
+```
+
+### Naming convention in this repo
+- Filenames follow the original pattern like: `ALWI_000.png`
+- **Image ↔ mask pairing rule:** the mask has the **exact same filename** as its image, but lives under the parallel `masks/` folder.
+
+### Notes about the `normal/` masks
+- The original dataset encodes background/normal tissue as **black** in the RGB mask.
+- In this repo, normal images are stored separately under `normal/images/`, and the corresponding `normal/masks/` are expected to be all-background (black) masks (verify visually if needed).
+
+---
+
+## 3) How to download / use (recommended)
+
+### A) Download only `BUS-UCLM/` from this HF repo
+```python
+from huggingface_hub import snapshot_download
+
+local_dir = snapshot_download(
+    repo_id="saiteja33/MedThink-Seg",
+    repo_type="dataset",
+    allow_patterns=["BUS-UCLM/**"],   # only pull BUS-UCLM subdir
+)
+print(local_dir)
+```
+
+### B) Build (image, mask) pairs from the folder structure
+```python
+from pathlib import Path
+
+root = Path(local_dir) / "BUS-UCLM"
+
+# lesion (benign+malignant)
+lesion_imgs = sorted((root / "images").glob("*.png"))
+lesion_pairs = [(p, (root / "masks" / p.name)) for p in lesion_imgs]
+
+# normal
+normal_imgs = sorted((root / "normal" / "images").glob("*.png"))
+normal_pairs = [(p, (root / "normal" / "masks" / p.name)) for p in normal_imgs]
+
+print("lesion pairs:", len(lesion_pairs))
+print("normal pairs:", len(normal_pairs))
+```
+
+### C) Load with 🤗 Datasets (two common options)
+
+**Option 1 — load just the images as an imagefolder dataset**
+```python
+from datasets import load_dataset
+
+ds_imgs = load_dataset("imagefolder", data_dir=str(root / "images"))
+ds_norm = load_dataset("imagefolder", data_dir=str(root / "normal" / "images"))
+
+print(ds_imgs)
+print(ds_norm)
+```
+
+**Option 2 — use `load_dataset` on the repo (works only if there is a dataset script/config)**
+```python
+from datasets import load_dataset
+ds = load_dataset("saiteja33/MedThink-Seg", data_dir="BUS-UCLM")
+print(ds)
+```
+
+---
+
+## Citation
+If you use BUS-UCLM, cite the Scientific Data paper and the Mendeley record:
+
+- Vallez, N. et al. *BUS-UCLM: Breast ultrasound lesion segmentation dataset.* Scientific Data (2025). DOI: `10.1038/s41597-025-04562-3`
+- BUS-UCLM (Mendeley Data, V3). DOI: `10.17632/7fvgj4jsp7.3`